Back-propagation neural network with enhanced neuron characteristics

A method for establishing a neural network comprising:

• producing an activation function and a first derivative of the activation function utilizing at least one neuron circuit in a feedforward network;
• providing updated weights to the feedforward network utilizing a weight updating circuit;
• receiving the first derivative of the activation function by an error back-propagation network;
• and
• providing weight change data information to the weight updating circuit from the error back-propagation network.

A neural network system comprising:

• a feedforward network comprising at least one neuron circuit for producing an activation function and a first derivative of the activation function, the at least one neuron circuit comprising:
• a linear resistor circuit comprising,
• a first linear resistor circuit transistor and a second linear resistor circuit transistor configured in a differential transistor pair, and
• a first reference voltage configured to cause the first linear resistor circuit transistor and the second linear resistor circuit transistor to operate in a linear range,
• a first differential circuit electrically connected to the linear resistor circuit, the first differential circuit configured to output a first differential circuit output, the first differential circuit comprises a first differential transistor pair having a first differential transistor pair first port electrically connected to a linear resistor circuit output and a first differential transistor pair second port electrically connected to a second reference voltage,
• a second differential circuit electrically connected to the linear resistor circuit, the second differential circuit configured to output a second differential circuit output, the second differential circuit comprises a second differential transistor pair having a second differential transistor pair first port electrically connected to the linear resistor circuit output and a second differential transistor pair second port electrically connected to a third reference voltage, the third reference voltage being less than the second reference voltage by a fixed amount, and
• wherein the first differential circuit output is the activation function and the first differential circuit output minus the second differential circuit output is the first derivative of the activation function, the error between a fitted sigmoid function and the first differential circuit output is less than 3%, the error between the first derivative of a fitted sigmoid function and the first differential circuit output minus the second differential circuit output is less than 5%, the first differential circuit output is shifted when the magnitude of the first reference voltage is varied, the linear resistor circuit is controllable by a first linear resistor circuit transistor gate voltage and a second linear resistor circuit transistor the gate voltage, a gain factor of the first differential circuit output is varied when the magnitude of the first linear resistor circuit transistor gate voltage and the second linear resistor circuit transistor the gate voltage are varied, and saturation levels of the activation function remain constant for different gain factor values;
• a weight updating circuit for producing updated weights to the feedforward network;
• an error back-propagation network for receiving the first derivative of the activation function and to provide weight change data information to the weight updating circuit;
• and
• wherein the neural network system is constructed on one electronic chip.